Deep learning generated CT-free PET with near-zero SUV bias in NSCLC

KEY POINTS

• The study used 100 patients with stage IIB/III NSCLC from the multicenter ACRIN 6668/RTOG 0235 dataset, all imaged with FDG-PET/CT before definitive chemoradiotherapy. Patients were split into 60 training, 10 validation and 30 independent test cases.
• A conventional 2D U-Net was trained to generate attenuation-corrected PET directly from non-attenuation-corrected emission PET, without using the CT attenuation map as an input. More than 15,000 paired axial images were available after preprocessing.
• Whole-image fidelity in the independent test cohort reached 36.2 dB PSNR and 0.967 SSIM. Within tumor regions, performance was slightly higher at 37.0 dB and 0.975, respectively.
• Lesion-level analysis included 42 NSCLC lesions in the 30 test patients. Mean SUVmax bias was only +0.72% ±3.04% (95% CI −0.41% to +1.85%; p=0.067), with no significant difference from standard CT-corrected PET.
• SUVmean showed a small systematic difference of −1.23% ±2.70% (95% CI −2.24% to −0.22%; p=0.005). Spatial overlap of automatically detected lesions remained high, with mean Dice 0.883 ±0.057 and HD95 of 3.37 ±1.29 mm.
• Image-derived physical characteristics were also preserved: contrast recovery was 99.2% ±2.1%, spatial resolution differed by only −1.5% from the CT-corrected reference (p=0.68), and contrast-to-noise ratio was 5.0% higher in the synthesized images.
• This is still a proof of concept rather than a CT-replacement study. It uses retrospective FDG scans from one historical trial dataset, a 2D architecture, and automated lesion identification based on an SUV >5 threshold; physical characteristics were assessed through image-derived surrogates rather than dedicated phantom validation, and prospective performance across scanners, protocols, tracers and patient populations remains unknown.

CLINICAL TAKEAWAY

SOURCE

Medical Dosimetry